Optical circuit switching gains traction as AI data centers scale

Optical circuit switching (OCS) is projected to reach $2.52 billion by 2032, spurred by exploding AI workloads that strain traditional electrical fabrics and by new solid‑state metasurface beamsteering that makes large‑scale OCS viable【1】.

Why AI data centers need OCS now

Traditional multi‑tier electrical switches (fat‑tree, Clos) add latency and power with every hop, a problem that grows as clusters move from hundreds to tens of thousands of GPUs【1】. AI training generates steady, high‑volume east‑west traffic, making each OEO conversion a measurable energy cost. By establishing dedicated optical paths, OCS eliminates repeated packet inspection, cuts latency, and reduces energy per bit—benefits that become decisive when networking power accounts for a sizable share of system budgets【1】.

New hardware pushes OCS into the mainstream

Salience Labs recently announced a 32‑port all‑optical silicon photonic switch, claiming it delivers the industry’s highest performance and can scale to 64‑ and 128‑port variants for larger AI clusters【2】. The design removes the need for separate optical transceivers, promising up to eight‑fold power savings versus conventional OEO solutions and latency reductions that can shave up to 80 % of tail latency【2】. These solid‑state devices avoid the mechanical complexity of earlier MEMS‑based OCS, offering higher reliability and the ability to integrate with existing transceivers【2】.

Market outlook and scaling potential

Analysts note that global data‑center capacity is expected to grow sixfold between 2025 and 2035, from 24.4 GW to 147.1 GW, driven by enterprise AI adoption and higher rack‑level power density【2】. This surge creates a pressing need for network fabrics that can handle massive, predictable data flows without excessive power draw. OCS architectures—ranging from compact 256 × 256 switches at the rack level to potential 10,000 × 10,000 backbones—are being explored to flatten network hierarchies and reduce tier count, directly addressing the emerging bottleneck of power‑hungry packet processing【1】.

What to watch

• Deployment milestones for Salience Labs’ 64‑ and 128‑port OCS products, slated for future releases.
• Power‑consumption benchmarks comparing OCS to conventional OEO switching as AI clusters exceed 10,000 GPUs.
• Industry adoption rates of solid‑state metasurface beamsteering in large‑scale data‑center fabrics.

The shift toward OCS reflects a broader architectural rethink: as AI workloads dominate traffic patterns, the network fabric must evolve from static electrical hierarchies to adaptable, energy‑efficient optical domains. The pace of hardware rollout and real‑world power savings will determine whether OCS becomes the new backbone of AI‑centric data centers.